Polylane

• Letting Polylane watch a Cloudflare Worker or Lambda function and open a pull request when a latency regression traces back to a recent code change
• Forwarding noisy Datadog or Sentry alerts so Polylane dismisses the false ones with reasons and investigates the rest
• Asking in Slack who owns a failing service, what changed, and whether the problem is on your side
• Requiring Polylane's production-impact check on pull requests, including ones opened by coding agents
• Giving Claude Code or Cursor production context over MCP before it changes an API or schema
• Having Polylane open pull requests that add missing logging and tracing to route handlers

Polylane is aimed at teams that want incidents to end in a reviewed pull request rather than a root-cause report, and its insistence on cited evidence and read-only defaults addresses the main worry about giving an agent production access. Its coverage of Cloudflare, Vercel, Fly.io, Railway, and Supabase fits startups on modern platforms better than enterprise operations stacks, which still need GCP, Azure, GitLab, and PagerDuty support. Because credit pools are stated as multiples of the Free allowance rather than in investigations, the practical way to judge cost is to connect one service on Free, compare its findings with a few recent incidents, and watch how quickly credits are used.

About Polylane

Polylane, built by Coreplane Labs in San Francisco and founded by Boris Tane (who founded Baselime, acquired by Cloudflare), treats operations as a closed loop. Connect cloud accounts, GitHub repositories, and observability providers, and it builds a live context graph of every resource, dependency, and repository, with change records for what appeared, changed, or disappeared. It reads metrics, logs, and traces on a schedule against each resource's own baselines, so there are no alert rules to write, and it triages alerts you forward from tools such as Datadog, Sentry, or Grafana Cloud. Each confirmed issue is worked by a single agent run that gathers evidence, tests hypotheses, and ends as resolved, diagnosed, or inconclusive. Every step cites the query, log line, or deploy behind it, and the run is a transcript you can watch, interrupt, or share. When the cause is a code defect in a connected repository, the same run writes the fix on a branch in a sandbox, checks it for risks such as locking migrations, and opens a pull request that your review and CI still gate. When code cannot fix it, Polylane escalates to an engineer with the root cause instead. It also reviews every pull request against the infrastructure it deploys to, passing by default and failing only when it can show how a line breaks production, answers on-call questions in Slack, and gives Claude Code, Cursor, Codex, and other MCP clients the same production context through one MCP server, with unlimited reads on every plan. Polylane's own engineering blog reports that, over the month around its switch from an 18-agent pipeline to one agent, median time from detection to pull request fell from 2.2 hours to 35 minutes and the share of detected issues ending in a pull request rose from 0.6% to 4.2%, so most runs still end as a report or a dismissal. Pricing is a flat monthly plan with a shared pool of cost-weighted AI credits, from a Free tier to $80, $200, and higher plans. The tradeoffs: it is in early access, repositories must be on GitHub, GCP, Azure, GitLab, and PagerDuty are not yet supported, the pricing page lists bring-your-own model key only on Enterprise, and running it in your own cloud is an Enterprise item that the public roadmap still marks as planned.